I checked the code and the query part starts from this place
query_engine_decompose = graph.as_query_engine(
custom_query_engines=custom_query_engines,
)
I could be wrong here but IMO, All the work happens in
ComposableGraphQueryEngine
,
Iterates over the user query against the set of query engines and then produce the results.
So for you to explore/debug further this would be where you should look:
https://github.com/run-llama/llama_index/blob/31d132c56b1836603d48e02786cd29a74a28f527/llama_index/query_engine/graph_query_engine.py#L50